fix: raise ValueError for unsupported MIME types in file_data URI path

[Raman369AI]

Relates to google#5022 and upstream PR google#5023

Problem

When a Part with file_data.file_uri has no determinable MIME type, the library fell back to application/octet-stream via _DEFAULT_MIME_TYPE. This value propagated to LiteLLM which raised a cryptic internal ValueError with no guidance for the user.

The same failure occurred when the caller explicitly set mime_type = "application/octet-stream" on the Part. Both cases reach the same failure point.

There was also an inconsistency between the two content paths:

• The inline_data path raises ValueError immediately for unsupported MIME types
• The file_data path silently used a fallback and failed later with a cryptic message

GcsArtifactService generates URIs like gs://bucket/artifact/0 with no extension and no MIME type, making ADK's own artifact system the primary trigger for this fallback.

Fix

Removes _DEFAULT_MIME_TYPE and raises ValueError early with an actionable message when the resolved MIME type is either unknown or application/octet-stream. This aligns the file_data path with the existing fail-fast behavior of the inline_data path.

The logic order is also corrected so providers that always produce a text fallback (anthropic, non-Gemini Vertex AI) and OpenAI/Azure HTTP media URLs are handled before the MIME type guard, keeping those paths unaffected.

Changes

• src/google/adk/models/lite_llm.py: remove _DEFAULT_MIME_TYPE, restructure file_uri handling block, raise ValueError for missing or generic MIME types
• tests/unittests/models/test_litellm.py: update two existing tests to assert the new ValueError, add one new test covering explicit application/octet-stream

Testing

pytest tests/unittests/models/test_litellm.py
241 passed, 5 errors (pre-existing, missing pytest-mock fixture)

Format verified with pyink. mypy comm -13 simulation: zero new errors.

[gemini-code-assist]

Code Review

This pull request updates the ADK to version 1.28.0, introducing significant security, concurrency, and stability enhancements. Key improvements include origin-check middleware for HTTP and WebSockets, strict agent name validation to prevent arbitrary imports, and the restriction of builder endpoints to web-enabled environments. The update also resolves concurrency races in session state management, adds regional support for Discovery Engine, ensures correct Pydantic required field parsing, provides consistent function call ID generation, and pins the LiteLLM dependency to a secure version. Feedback was provided regarding the risk of silent data loss when configuring models to ignore extra fields, suggesting a more explicit approach to unknown data.

[gemini-code-assist]

Changing extra='forbid' to extra='ignore' in model_config can lead to silent data loss if external data sources introduce new fields that are not explicitly handled by the model. While it can improve compatibility, it's important to be aware of the trade-offs. Consider if extra='allow' or a more explicit handling of unknown fields would be more appropriate depending on the expected data evolution.